Ditto CEO Accuses AI Labs of Hoarding Reasoning Tokens

Ditto CEO Adam Fish alleges that AI model providers OpenAI and Anthropic utilize legal ambiguities in their service agreements to retain ownership of reasoning tokens. These tokens represent the intermediate computational steps a model performs before delivering a final answer. Fish argues that while these companies promise not to train on Customer Content—defined as Input and Output—the intermediate reasoning data falls outside those definitions, allowing providers to maintain a digital scratchpad of derived business insights.

Fish notes that OpenAI explicitly hides raw chains of thought to maintain a competitive advantage, while Anthropic bills customers for reasoning tokens that remain encrypted and invisible to the user. He suggests this data is strategically valuable for model distillation and may be used for training despite general no-training promises. These claims align with previous warnings from Palantir CEO Alex Karp regarding data ownership and caching. Fish identifies OpenAI CEO Sam Altman and Anthropic CEO Dario Amodei as the leaders capable of resolving these contractual ambiguities.